[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Pulp-list] Sanity Check

On 09/23/2010 12:15 PM, Bryan Kearney wrote:
On 09/23/2010 12:58 PM, Jeff Ortel wrote:

On 09/23/2010 11:08 AM, Bryan Kearney wrote:
On 09/23/2010 11:58 AM, Jeff Ortel wrote:

On 09/23/2010 10:49 AM, Bryan Kearney wrote:
On 09/23/2010 09:58 AM, Jeff Ortel wrote:

I'd like to get a quick sanity check on how I designed the
of API methods (that rely on RMI on the agent) through the task
subsystem. The approach is simple enough. Hopefully, the long
write up
will not make it seem more complicated than it really is :)

First, add classes to pulp.server.async.py that mirror the
pulp.server.agent.Agent class but performs RMI asynchronously by
and specifies the same correlation tag as the reply listener. The
pulp.server.agent.Agent does synchronous RMI by default and this way,
the replies are guaranteed to match what the reply listener is

Second, the ReplyListener is added to pulp.server.async.py. It's
job is
to listen for asynchronous RMI replies based on the same correlation
specified when the RMI was invoked as described above. When replies
received, the associated task is updated by calling either
Task.succeeded() or Task.failed(). The task is also updated with the
returned value or exception as appropriate.

Third, I extended the pulp.server.tasking.task.Task as AsyncTask.
its superclass, it expects its callable to be asynchronous. Further,
when the task runs the callable, the task's state is not advanced in
anticipation that its succeeded() or failed() methods will be
called by
an external event.

Last, and most significant, is how the API classes leverage (1-3).
API methods such as ConsumerApi.installpackages() need to do
on the agent and then update consumer history if the agent (RMI) was
successful. This is what makes this kind of tricky. So, what I
to implement this by subclassing the AsyncTask as
pulp.server.api.consumer.InstallPackages. This specialized task is
created/enqueued and returned by ConsumerApi.installpackages() giving
InstallPackages.install() as the task's target callable. When
it sends the RMI. Then, when the reply comes in and
Task.succeeded() is
called by the queue. This method is overridden in InstallPackages
calls super.succeeded() and then updates the consumer history.

It looks like this:


[ task runs Task.callable() ]
[ RMI to agent ]
[ agent reply received on QPID queue ]

One follow up question.. what happens when ConsumerApi.installpackages()
is version 2 and InstallPackages.install() is version 1?

They both live on the server and would be versioned together.

Are we following the RHQ model where we deliver outside of RPMS?

Ah, ConsumerApi.installpackages() both are on the server InstallPackages.install() but I assume you're asking about API version mismatches between the server and client, right?

No, are not following the RHQ model and updating the client (agent) outside of RPMs.

The problem of API version mismatch is classic. The key, IMHO, is to bound the problem. No amount of messaging magic can make arbitrary API mismatches go away. So, we need to solve this with a combination of policy and technology. Well, mostly policy.

Here's an initial stab at policy:
 * We should support a client that is no more than 1 major version behind the server.
    Any more and the client must be updated (install newer rpm).
 * APIs must be backwards compatible for at least major 1 version back.
 * New APIs need to be added to the client and released 1 major version ahead of
    being used on the server.
 * Client should squeal or gently notify admins when client API version trails
    the server.  The client should report this to the server so we can
    display in UIs.

The "1 major version" may be too course grained but it's a place to start. We can adjust as needed. Suggestions?


Is the assumption that the task is stored in either the DB or a user

Currently, tasks are persisted using an in-memory store. Eventually, we
need to add a non-volatile store so tasks are preserved across


Also... how does this patern change if I want to install
packages at noon next tuesday?

The messaging and agent frameworks already support the concept of
execution (or maintenance) "windows" for RMI calls. That is,
asynchronous RMI with a date/time window specified.

So, we have two choices:

1) Dispatch though the task subsystem as defined here but include a
parameter to
specify the window as part of the RMI call to the agent.

2) Introduce task scheduling into the task subsystem.

So this would allow storing the task locally? One feedback we got was to
not have the task execution dependent upon "the next time I poll"

In #1, the RMI request is stored on the client. When the current time is
within the specified window, the RMI is executed. No communication (or
polling) with the server is required. Is that what you're asking?

Yes, I believe the model here is that in todays world the client will
execute "The closest time it wakes up to the execution time, and then
polls" which is slightly different.

-- bk

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]